Training Artificial Neural Networks Using Context-Dependent Gating with Weight Stabilization

A computer-implemented method, carried out by a computing device, for computationally training an artificial neural network (ANN) implemented in the computing device, the method comprising:

• at the computing device, training the ANN to predict ground truth data for a first type of task using as input a first set of training associated with the first type of task, wherein during training, values of only those weights associated with a first subset of neurons of the ANN are adjusted, and wherein the first subset of neurons is selected based on an identity of the first type of task;
• at the computing device, training the ANN to predict ground truth data for a second type of task using as input a second set of training associated with a second type of task, wherein the second type of task is different from the first type of task, wherein during training, values of only those weights associated with a second subset of neurons of the ANN are adjusted, and wherein the second subset of neurons is selected based on an identity of the second type of task;
• and
• during training of the ANN for any given type of task, adjusting of the value of any given weight associated with neurons of the ANN subject to a bias in dependence of whether or not the ANN has been previously trained for one or more task types different from the given type.

A computing device comprising:

• one or more processors;
• and
• memory configured to store computer-executable instructions that, when executed by the one or more processors, cause the computing device to carry out operations including:
• applying an artificial neural network (ANN) implemented in the computing device to the data associated with the first type of task to predict output data for the first type of task, wherein only a first subset of neurons of the ANN are activated when applying the ANN to the data associated with the first type of task;
• receiving data associated with a second type of task, wherein the second type of task is different from the first type of task;
• and
• applying the ANN to the data associated with the second type of task to predict output data for the second type of task, wherein only a second subset of neurons of the ANN are activated when applying the ANN to the data associated with the second type of task,
• wherein prior to receiving the data associated with the first and second types of tasks, the ANN is trained to:
• (i) predict ground truth data for the first type of task using as input a first set of training associated with the first type of task, wherein during training, values of only those weights associated with the first subset of neurons of the ANN are adjusted, and wherein the first subset of neurons is selected based on an identity of the first type of task, and
• (ii) predict ground truth data for the second type of task using as input a second set of training associated with the second type of task, wherein during training, values of only those weights associated with a second subset of neurons of the ANN are adjusted, and wherein the second subset of neurons is selected based on an identity of the second type of task,
• and wherein, during training of the ANN for any given type of task, adjusting of the value of any given weight associated with neurons of the ANN is subject to a bias in dependence of whether or not the ANN has been previously trained for one or more task types different from the given type.